Wavelet Thresholding-Based Despeckling of COSMO-SkyMed SAR Image

Abstract

The selection of an efficient speckle filter for SAR imagery primarily depends upon a specific application of interest and statistical characteristics of the noise present in SAR datasets. The main goal of this study is to assess the performance of the two wavelet shrinkage-based filtering techniques (VISU shrink and SURE shrink) against two spatial adaptive filters (Enhanced Lee and Gamma MAP) and one non-local filter (NL-SAR) for the removal of speckle noise from high-resolution COSMO-SkyMed (CSK) SAR datasets. Before applying these filters to real CSK datasets, they are tested on synthetically generated speckled test datasets and benchmark simulated SAR datasets. Experimental analysis has been conducted on synthetically generated speckled datasets based on varying level of speckle noise introduced on test images. In case of benchmark datasets, numerous qualitative and quantitative measures are observed and evaluated. To find the best filter for real CSK data, a Pareto optimality concept has been used where the coefficient of variation is the parameter considered. From the findings, it is evident that VISU shrink-generated speckle filtering solution is non-dominated by all the other filtering solutions except NL-SAR-based speckle suppression in smooth areas. Considering the various user-defined situations of homogeneity and heterogeneity in the SAR scene, an overall performance index is formulated and VISU shrink performs the best in all user-defined conditions.